Artificial Intelligence: Twelve Days of Problems

Only two weeks since my last AI post, but a lot has happened – and it isn’t good.

I don’t deal on speculation and rumors any more than future-only “progress,” so you won’t read here about the growing set of people thinking that Nvidia stock is heading for a steep decline – as with AI’s projected, expected, touted world-beating capabilities and applications, we’ll deal with it when it happens.  But here are harder concerns.

In “The age of AI BS” (Business Insider, March 27^th), Emily Stewart told us about “”AI washing,” or companies giving off a false impression that they’re using AI so they can amp up investors,” which has precipitated formal charges and legal settlements, as well as ChatGPT’s marketing campaign being mainly “to raise money, attract talent, and compete in the hypercompetitive tech industry,” a situation where “overselling has become a near-constant of the AI landscape,” and “it’s not super clear what the present capabilities of the technology even are, let alone what they might be in the future, so making bold, concrete claims about the way it’s going to affect society seems presumptuous.”  A “senior industry analyst” said that “a lot of these companies are not yet showcasing exactly what type of revenue they’re getting from AI yet because it’s still so small.” Stewart ended with “anyone who tells you they know exactly what is going on in AI and where it’s headed is lying.”

In response to a good and timely question, Zvi Mowshowitz informed us, in the New York Times on March 28^th, “How A.I. Chatbots Become Political.”  Although “our A.I. systems are still largely inscrutable black boxes” according to political-view assessment tests, most lean liberal and to some extent libertarian.  The biases may have been introduced during “fine tuning,” when technicians adjust outcomes, and from earlier developments, “because models learn from correlations and biases in training data, overrepresenting the statistically most likely results.”  There are now three versions of one major product, LeftWingGPT and RightWingGPT, which were evaluated as matching their names, and DepolarizingGPT, which still skewed slightly to the libertarian left.  A cause for concern, as “we may have individually customized A.I.’s telling us what we want to hear,” which may not end up being constructive.  This area, still, looks under control, but all should be aware of the biases these tools all carry.

Could it be, already, that “A.I.-Generated Garbage Is Polluting Our Culture” (Erik Hoel, The New York Times, March 31^st)?  In places, anyway.  One the author documented is recent academic peer reviews, which showed expressions known to be “the favorite buzzwords of modern large language models like ChatGPT” 10 to 34 times as often as in 2022.  Another example is nonsensical and uncorrected videos for children.  That sort of thing could cause a state “when future A.I.’s are trained, the previous A.I. output will leak into the training set, leading to a future of copies of copies of copies, as content (becomes) more stereotypical and predictable,” called “model collapse.”

Moving on, “It looks like it could be the end of the AI hype cycle” (Hasan Chowdhury, Business Insider, April 3^rd).  Not yet, though with the amount of accumulating doubt that may happen soon.  Along with sky-high importance assessments from Bill Gates and Elon Musk, Gary Marcus, who testified to the Senate on the technology, “predicted the generative AI bubble could burst within the next 12 months,” noting that “the industry is spending much more money than it’s raking in,” a great deal given that, per Crunchbase, “generative AI and AI-related startups raised almost $50 billion last year.”  Overall, “the verdict is still out on whether the companies behind foundation AI models dependent on expensive chips can turn their products into viable, profitable businesses.”

Back to a problem mentioned above, “Big Tech needs to get creative as it runs out of data to train its AI models.  Here are some of its wildest solutions” (Lakshmi Varanasi, Business Insider, April 7^th).  If, “according to Epoch, an AI research institute,” less than three years from now “all the high-quality data could be exhausted.”  Companies will need to do something, which could include “tapping consumer data available in Google Docs, Sheets, and Slides” which Google has already contemplated; buying an entire large publisher, such as Simon & Schuster, for its copyrighted information; “generating synthetic data,” created by AI modules themselves, using speech recognition to tap YouTube videos; and incorporating pictures from Photobucket, which hosted those from former large social media sites Myspace and Friendster.  In the meantime, we saw “How Tech Giants Cut Corners to Harvest Data for A.I.” (Cade Metz et al., The New York Times, April 6^th).  One way was the same YouTube idea, which was input into GPT-4 – this piece also mentioned the solutions Varanasi revealed. 

On the positive AI side, all I have seen these past two weeks is incomplete and future-bound.  For its actual advancement, they were a loser for artificial intelligence.  Will it get better, or worse?  I will keep you informed.